InfoResearchPreprint
As Student Rollouts Continue, Visual Backdoors Keep Growing
- Published
- Record updated
Summary
This paper studies how visual backdoors propagate from teacher to student models during On-Policy Distillation (OPD) for Vision-Language Models (VLMs). Experiments across multiple model families and datasets, plus theoretical analysis, show backdoors rarely emerge at low poisoning ratios but reach nearly 100% attack success rate once a critical threshold is exceeded. The authors attribute propagation to the continuous accumulation of anomalous probability mass from the teacher, and propose a trigger design that enables propagation with zero explicit trigger injection.
Related items
- HighCVE-2026-77283: obsidian-sync-mcp HTTP /mcp endpoint DNS rebinding and cross-origin vault accessSimilar attack · NVD/CVE Database
- HighCVE-2026-75612: yt-mcp arbitrary file read through add_attachment file_pathSimilar attack · NVD/CVE Database
- MediumCVE-2026-75518: DOMShell MCP server eval tool bypasses write permission checksSimilar attack · NVD/CVE Database
- HighCVE-2026-62964: Gortex arbitrary file write through MCP file toolsSimilar attack · NVD/CVE Database
- HighCVE-2026-127222: SGLang missing authorization on POST /load_lora_adapter_from_tensorsSimilar attack · NVD/CVE Database